CN-122017613-A - Multi-type fault diagnosis method for lithium iron phosphate battery based on reinforcement learning

CN122017613ACN 122017613 ACN122017613 ACN 122017613ACN-122017613-A

Abstract

The invention discloses a multi-type fault diagnosis method for a lithium iron phosphate battery based on reinforcement learning, which comprises the following steps of collecting multi-source data in the running process of the lithium iron phosphate battery, preprocessing the multi-source data, generating candidate fault hypotheses according to trigger observation results, constructing corresponding short-time response prediction paths, generating prediction response sequences, collecting newly-increased running data based on the trigger observation results, extracting the newly-increased response sequences, reserving, degrading or eliminating the candidate fault hypotheses, inputting the candidate shrinkage results into a reinforcement learning observation action selection strategy, updating the candidate shrinkage results, judging whether convergence conditions are met according to the candidate shrinkage results, outputting final fault diagnosis results, and outputting fault types, risk grades and corresponding diagnosis bases according to the final fault diagnosis results. The invention adopts reinforcement learning to realize multi-fault diagnosis of the lithium iron phosphate battery.

Inventors

FANG HUAMING
ZHANG HUITING
LIU YAPING
SHAO JIZHONG

Assignees

江苏东润动力科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260401

Claims (10)

1. A multi-type fault diagnosis method of a lithium iron phosphate battery based on reinforcement learning is characterized by comprising the following steps: The method comprises the steps of collecting multi-source data in the running process of the lithium iron phosphate battery, preprocessing the multi-source data, and identifying abnormal triggering events to obtain a triggering observation result; Generating candidate fault assumptions according to the triggering observation result to obtain a candidate fault assumption set; Respectively constructing corresponding short-time response prediction paths aiming at each candidate fault hypothesis in the candidate fault hypothesis set, and generating a prediction response sequence to obtain a candidate prediction result set; Acquiring newly-added operation data based on a trigger observation result, extracting newly-added response sequences, and respectively carrying out consistency comparison with the predicted response sequences to obtain a candidate consistency result set; According to the candidate consistency result set, carrying out retaining, degrading or eliminating treatment on the candidate fault hypothesis, and updating the diagnosis priority of the rest candidate fault hypothesis to obtain a candidate shrinkage result; inputting the candidate shrinkage results into a reinforcement learning observation action selection strategy to obtain an observation action selection result when the number of residual candidate fault assumptions in the candidate shrinkage results is more than one, and updating the candidate shrinkage results according to the observation action selection result; judging whether convergence conditions are met or not according to the candidate contraction results, obtaining convergence judgment results, and outputting final fault diagnosis results based on the convergence judgment results; And outputting the fault type, the risk level and the corresponding diagnosis basis according to the final fault diagnosis result to obtain a fault diagnosis output result.
2. The reinforcement learning-based multi-type fault diagnosis method for a lithium iron phosphate battery according to claim 1, wherein the preprocessing includes denoising, outlier rejection, time alignment and normalization.
3. The method for diagnosing multiple types of faults of the lithium iron phosphate battery based on reinforcement learning according to claim 1, wherein the generating candidate fault hypotheses according to the trigger observation result, and obtaining a candidate fault hypothesis set specifically comprises: extracting trigger discriminants corresponding to the current abnormal state from the trigger observation result to obtain a trigger discriminant set; Respectively calculating the internal short circuit fault assumption support, the thermal abnormal fault assumption support, the capacity attenuation fault assumption support and the consistency abnormal fault assumption support according to the trigger discriminant set to obtain an assumption support set; normalizing each support degree in the hypothesized support degree set to obtain a normalized support degree set; sequencing each candidate fault hypothesis according to the size of each normalized support degree in the normalized support degree set to obtain a candidate hypothesis sequencing result; and selecting candidate fault hypotheses according to the candidate hypothesis ranking result, and organizing each candidate fault hypothesis into a candidate fault hypothesis set.
4. The method for diagnosing multiple types of faults of the lithium iron phosphate battery based on reinforcement learning according to claim 1, wherein the steps of constructing corresponding short-time response prediction paths for each candidate fault hypothesis in the candidate fault hypothesis set and generating a prediction response sequence respectively, and obtaining the candidate prediction result set specifically comprise: For each candidate fault hypothesis in the candidate fault hypothesis set, reading the single voltage, the single temperature, the voltage difference between the single cells and the temperature difference between the single cells corresponding to the current sampling time as initial response values; generating predicted variation of each predicted moment according to the fault evolution coefficient set to obtain a predicted variation set; generating a voltage change sequence corresponding to each candidate fault hypothesis according to the predicted change quantity set; generating a temperature change sequence corresponding to each candidate fault hypothesis according to the predicted change quantity set; Generating a monomer deviation change sequence corresponding to each candidate fault hypothesis according to the voltage change sequence and the temperature change sequence; according to each candidate fault hypothesis type in the candidate fault hypothesis set, carrying out path combination on the voltage change sequence, the temperature change sequence and the monomer deviation change sequence to obtain a short-time response prediction path; and generating a corresponding prediction response sequence according to each short-time response prediction path, and summarizing according to the candidate fault hypothesis types to obtain a candidate prediction result set.
5. The reinforcement learning-based lithium iron phosphate battery multi-type fault diagnosis method according to claim 4, wherein the set of fault evolution coefficients comprises: Collecting historical operation data of the lithium iron phosphate battery under different fault types, and classifying the historical operation data according to the fault types to obtain a historical fault sample set; Extracting voltage variation, temperature variation and deviation variation of a plurality of continuous sampling moments after corresponding faults occur from various historical fault samples in a historical fault sample set to obtain a historical evolution feature set; aiming at various historical fault samples in the historical evolution feature set, respectively calculating the average voltage change rate, the average temperature change rate and the average deviation change rate under the corresponding fault types to obtain a fault evolution statistical result; And respectively determining a voltage evolution coefficient, a temperature evolution coefficient and a deviation evolution coefficient corresponding to each fault type according to the fault evolution statistical result to obtain a fault evolution coefficient set.
6. The method for diagnosing multiple types of faults of the lithium iron phosphate battery based on reinforcement learning according to claim 1, wherein the steps of collecting new operation data based on triggering observation results, extracting new response sequences, and respectively comparing the new response sequences with predicted response sequences in a consistent manner to obtain candidate consistency result sets specifically comprise: after triggering the current sampling time corresponding to the observation result, continuously collecting newly-added operation data at a plurality of continuous sampling times to obtain a newly-added operation data sequence; Extracting a new voltage response sequence, a new temperature response sequence and a new monomer deviation response sequence corresponding to the candidate prediction result set from the new operation data sequence to obtain a new response sequence set; aiming at a prediction response sequence corresponding to any candidate fault hypothesis in the candidate prediction result set, respectively calculating voltage consistency deviation, temperature consistency deviation, voltage difference consistency deviation and temperature difference consistency deviation to obtain a subentry consistency deviation result; Calculating the comprehensive consistency deviation of any candidate fault hypothesis at a plurality of continuous newly-increased sampling moments according to the sub-term consistency deviation result to obtain a comprehensive consistency deviation result; calculating a consistency score corresponding to any candidate fault hypothesis according to the comprehensive consistency deviation result to obtain a consistency score result; And summarizing consistency scores corresponding to the candidate fault hypotheses in the candidate fault hypothesis set to obtain a candidate consistency result set.
7. The method for diagnosing multiple types of faults of the lithium iron phosphate battery based on reinforcement learning according to claim 1, wherein the steps of reserving, degrading or eliminating candidate fault hypotheses according to the candidate consistency result set and updating the diagnosis priority of the remaining candidate fault hypotheses to obtain candidate shrinkage results specifically comprise: reading the consistency score corresponding to each candidate fault hypothesis in the candidate fault hypothesis set from the candidate consistency result set to obtain a candidate score set; Aiming at each consistency score in the candidate score set, calculating the score variation of each candidate fault hypothesis in a plurality of continuous diagnosis periods by combining the consistency scores of the corresponding candidate fault hypotheses in the previous diagnosis period to obtain a score variation result; According to the candidate score set and the score change result, carrying out retention processing, degradation processing or elimination processing on each candidate fault hypothesis in the candidate fault hypothesis set to obtain a candidate processing result; Assigning a maintenance diagnosis priority to the candidate fault hypothesis for performing the retention process, reducing the diagnosis priority to the candidate fault hypothesis for performing the degradation process, and deleting the candidate fault hypothesis for performing the elimination process from the candidate fault hypothesis set to obtain a priority adjustment result; reordering the candidate fault hypothesis after the retention processing and the candidate fault hypothesis after the degradation processing according to each diagnosis priority in the priority adjustment result to obtain a candidate ordering update result; And forming a residual candidate fault hypothesis set according to the candidate sorting updating result and the deleting result corresponding to the eliminating process to obtain a candidate shrinkage result.
8. The method for diagnosing multiple types of faults of a lithium iron phosphate battery based on reinforcement learning according to claim 1, wherein the steps of inputting the candidate shrinkage result into a reinforcement learning observation action selection strategy to obtain an observation action selection result, and updating the candidate shrinkage result according to the observation action selection result specifically comprise: the number of residual candidate fault hypotheses in the candidate shrinkage results is greater than one, and a current observation decision state is formed based on the candidate shrinkage results, so that an observation decision state result is obtained; Inputting the observation decision state result into a reinforcement learning observation action selection strategy, and calculating action values corresponding to a transient voltage drop observation action, a voltage recovery observation action, a temperature-voltage coupling observation action and a monomer deviation expansion observation action to obtain an action value set; Selecting the observation action with the largest action value as the next round of observation action according to the action value set to obtain an observation action selection result; Acquiring new operation data corresponding to the observation action according to the observation action selection result to obtain a new operation data sequence corresponding to the action; Extracting a corresponding newly added response sequence according to the newly added operation data sequence corresponding to the action, and updating the candidate consistency result set to obtain an updated candidate consistency result set; and re-executing the retention process, the degradation process or the elimination process of each candidate fault hypothesis in the candidate fault hypothesis set according to the updated candidate consistency result set, and updating the diagnosis priority of the rest candidate fault hypotheses to obtain updated candidate contraction results.
9. The reinforcement learning-based multi-type fault diagnosis method for lithium iron phosphate battery according to claim 1, wherein the determining whether the convergence condition is satisfied according to the candidate shrinkage result to obtain a convergence determination result, and outputting a final fault diagnosis result based on the convergence determination result specifically comprises: reading a residual candidate fault hypothesis set, diagnosis priorities corresponding to all residual candidate fault hypotheses, consistency scores corresponding to all residual candidate fault hypotheses and fault types corresponding to all residual candidate fault hypotheses in the candidate shrinkage results to obtain a convergence judgment input result; counting the number of candidate fault hypotheses in the residual candidate fault hypothesis set according to the convergence judgment input result to obtain a candidate number judgment result; Identifying high-risk candidate fault hypotheses according to the convergence judging input result, and counting the continuous period number of each high-risk candidate fault hypothesis for maintaining the highest diagnosis priority in a plurality of continuous diagnosis periods to obtain a high-risk priority judging result; counting whether the rest candidate fault hypotheses in the candidate fault hypothesis set are all eliminated according to the convergence judging input result to obtain an elimination state judging result; Carrying out convergence condition judgment according to the candidate number judgment result, the high-risk priority judgment result and the elimination state judgment result to obtain a convergence judgment result; and when the convergence judging result meets the convergence condition, reading the fault type corresponding to the candidate fault hypothesis with the highest diagnosis priority in the rest candidate fault hypothesis set, and outputting the fault type as a final fault diagnosis result.
10. The method for diagnosing multiple types of faults of the lithium iron phosphate battery based on reinforcement learning according to claim 1, wherein the outputting of the fault type, the risk level and the corresponding diagnosis basis according to the final fault diagnosis result, the obtaining of the fault diagnosis output result specifically comprises: The type of the candidate fault hypothesis corresponding to the final fault diagnosis result, the diagnosis priority of the corresponding candidate fault hypothesis in the candidate contraction result, the consistency score of the corresponding candidate fault hypothesis in the candidate consistency result set and the meeting condition of the corresponding convergence condition in the convergence judgment result are read to obtain output judgment information; Determining the fault type according to the output judging information to obtain a fault type output result; calculating a risk grade evaluation value corresponding to the final fault diagnosis result according to the output judgment information to obtain a risk evaluation result; Determining a risk level according to the risk evaluation result to obtain a risk level output result; extracting corresponding diagnosis basis according to the final fault diagnosis result, the candidate prediction result set, the candidate consistency result set and the convergence judgment result to obtain a diagnosis basis output result; And summarizing the fault type output result, the risk level output result and the diagnosis basis output result to obtain a fault diagnosis output result.

Description

Multi-type fault diagnosis method for lithium iron phosphate battery based on reinforcement learning Technical Field The invention relates to the technical field of battery fault diagnosis, in particular to a multi-type fault diagnosis method for a lithium iron phosphate battery based on reinforcement learning. Background The existing fault diagnosis method of the lithium iron phosphate battery generally identifies abnormal states of the battery based on operation data such as voltage, current, temperature, state of charge and state of health. In the prior art, one type of method adopts threshold judgment, rule matching or state parameter analysis to judge internal short circuit, thermal abnormality, capacity attenuation and consistency abnormality, the other type of method carries out feature extraction and classification recognition on collected time sequence data by constructing a data model, and the other type of method utilizes the difference between a predicted value and an actual value to detect and early warn the abnormality. In the prior art, fault identification is mostly directly performed based on current observation data, or comparison is performed only according to a single prediction result and an actual result, and when multiple types of faults coexist or the fault performances are close to each other, it is difficult to gradually distinguish multiple candidate faults. Meanwhile, the selection of the follow-up observation content by the existing method generally depends on a fixed flow, and a mechanism for dynamically adjusting the observation action according to the change condition of the candidate fault is lacking, so that the candidate fault in the fault diagnosis process has insufficient shrinkage capability, and the pertinence of fault type judgment and the integrity of diagnosis output are finally influenced. Therefore, how to provide a multi-type fault diagnosis method for lithium iron phosphate batteries based on reinforcement learning is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a multi-type fault diagnosis method for a lithium iron phosphate battery based on reinforcement learning, which comprehensively utilizes multi-source operation data analysis, candidate fault hypothesis generation, short-time response prediction, consistency comparison and reinforcement learning observation action selection methods, and details the implementation process of gradual shrinkage diagnosis and convergence judgment on internal short-circuit faults, thermal anomaly faults, capacity attenuation faults and consistency anomaly faults. According to the embodiment of the invention, the multi-type fault diagnosis method for the lithium iron phosphate battery based on reinforcement learning comprises the following steps of: The method comprises the steps of collecting multi-source data in the running process of the lithium iron phosphate battery, preprocessing the multi-source data, and identifying abnormal triggering events to obtain a triggering observation result; Generating candidate fault assumptions according to the triggering observation result to obtain a candidate fault assumption set; Respectively constructing corresponding short-time response prediction paths aiming at each candidate fault hypothesis in the candidate fault hypothesis set, and generating a prediction response sequence to obtain a candidate prediction result set; Acquiring newly-added operation data based on a trigger observation result, extracting newly-added response sequences, and respectively carrying out consistency comparison with the predicted response sequences to obtain a candidate consistency result set; According to the candidate consistency result set, carrying out retaining, degrading or eliminating treatment on the candidate fault hypothesis, and updating the diagnosis priority of the rest candidate fault hypothesis to obtain a candidate shrinkage result; inputting the candidate shrinkage results into a reinforcement learning observation action selection strategy to obtain an observation action selection result when the number of residual candidate fault assumptions in the candidate shrinkage results is more than one, and updating the candidate shrinkage results according to the observation action selection result; judging whether convergence conditions are met or not according to the candidate contraction results, obtaining convergence judgment results, and outputting final fault diagnosis results based on the convergence judgment results; And outputting the fault type, the risk level and the corresponding diagnosis basis according to the final fault diagnosis result to obtain a fault diagnosis output result. Optionally, the preprocessing includes denoising, outlier rejection, time alignment and normalization. Optionally, generating the candidate fault hypothesis according to the trigger observation result, and obtaining the candidate fault